A Benchmark for Replay Methods in Lifelong Learning

Abstract

With the capacity of lifelong learning, humans can continuously acquire knowledge throughout their lifespan. However, computational systems are not, in general, capable of learning tasks sequentially. This long-standing challenge for machine learning and neural networks is called catastrophic forgetting. Multiple solutions (regularization, architectural, and rehearsal methods) have been proposed to overcome this limitation, but problems remain. On the one hand, few existing datasets faithfully reflect the real-world's varying environmental conditions. On the other hand, limited benchmarks are available to evaluate and compare the emerging techniques.

In this project, I first construct a lifelong robotic vision dataset OpenLORIS-Object, isolating varying factors in robotic vision under real-world scenarios (illumination, occlusion, distance, and clutter) and defining their difficulty levels explicitly. Then, I make an in-depth evaluation of the memory replay methods, exploring: the efficiency, accuracy and, scalability of different sampling strategies (Random, Confidence, Entropy, Margin, K-means, Core-Set, Maximally Interfered Retrieval, and Bayesian Disagreement) when selecting replay data; the relation between the difficulty of replaying data and the efficiency; and the difference between Experience replay and Generative replay. All the experiments are conducted with multiple datasets under various domains (MNIST, CIFAR-10, MiniImagenet, OpenLORIS-Object). Finally, I will provide a practical solution for selecting replay methods for various data distributions in conclusion.